王妍玮, 张佳宇, 陈凯云,等.将异步事件流转换为网格表示的方法研究[J].河南理工大学学报（自然科学版）,2025,44(5):17-26.

WANG Y W, ZHANG J Y, CHEN K Y,et al.Study on the method of converting asynchronous event stream into grid representation[J].Journal of Henan Polytechnic University(Natural Science) ,2025,44(5):17-26.

将异步事件流转换为网格表示的方法研究

王妍玮, 张佳宇, 陈凯云, 任春平

黑龙江科技大学 机械工程学院，黑龙江 哈尔滨 150022

摘要: 目的 针对异步事件流的复杂性及稀疏性带来的数据分析复杂化和存储计算效率低等问题，提出一种将异步事件流转换为网格表示的方法。  方法 首先采用Dirac脉冲以函数形式替代每一个事件，并将其表示为一个事件场集合，根据事件的张量特性为每个丢失同一类信息最多的事件分配一个该类别的平均测量值，减小计算量的同时保留事件的高动态分辨率；其次，选择可直接利用的数据，寻找最佳测量函数候选项的多层感知器MLP代替原有手动选择的聚合核函数，在ECTResNet中卷积，通过定期采样降低维度并保留关键信息进行量化处理；然后，经过卷积将事件网格化，通过连续的三维空间离散产生一个固定大小的网格；最后，将网格化的事件流转换为可深度学习的网格表示形式输出。 结果 把所提方法在数据集N-Cars和N-Caltech101上进行分析，经网格转换后的输出表示识别准确率可达97.07%和87.72%，比事件尖峰张量方法分别提高了10.09%和11.44%。  结论 实验表明，将异步事件流转换为网格表示能更好地适应深度学习模型，提高事件处理和识别的准确性和效率，并支持端到端的表示学习，可广泛应用于传感器数据处理与事件识别领域。

关键词:事件相机;异步事件流;深度学习;卷积神经网络

DOI:10.16186/j.cnki.1673-9787.2024070033

基金项目:国家自然科学基金资助项目（52204131）；黑龙江省重点研发计划战略研究专项项目（GA23A910）；黑龙江省科研基本业务费项目（2022-KYYWF-0527）

收稿日期:2024/07/05

修回日期:2024/10/11

出版日期:2025/07/23

Study on the method of converting asynchronous event stream into grid representation

Wang Yanwei, Zhang Jiayu, Chen Kaiyun, Ren Chunping

College of Mechanical Engineering， Heilongjiang University of Science & Technology， Harbin 150022， Heilongjiang， China

Abstract: Objectives To address the complexity and sparsity of asynchronous event streams， which complicate data analysis， reduce storage and computational efficiency， a method was proposed to convert asynchronous event stream into grid representation.  Methods Each event was replaced by a Dirac delta function and represented as a set of event fields. Based on tensor characteristics， average measurements were assigned to events missing the same category of information， reducing computation while preserving high dynamic resolution. Usable data were selected， and a multilayer perceptron （MLP） was used to replace manually chosen aggregation kernels to identify optimal measurement functions. In ECTResNet， convolution was performed and dimension was reduced through periodic sampling to retain key information for quantization. The convolved data were discretized in continuous 3D space to generate a fixed-size grid. Finally， the event stream was transformed into a grid representation suitable for deep learning. Results The proposed method was evaluated on the N-Cars and N-Caltech101 datasets. Recognition accuracies reached 97.07% and 87.72%， respectively， improving by 10.09% and 11.44% over the event spike tensor method.  Conclusions Experiments showed that converting asynchronous event stream into grid representation enhanced compatibility with deep learning models， improved accuracy and efficiency of event processing and recognition， and enabled end-to-end representation learning. This approach held broad potential in sensor data processing and event recognition.

Key words:event camera;asynchronous event stream;deep learning;convolutional neural network